openspec/changes/archive/2026-09-10-add-cartridge-model/tasks.md

Tasks

  • Advise reader (other family) on this architecture change (send-back 2026-09-10; three spec pins amended)
  • Catalog schema: id, kind, shelf, exclusive tags, costs per box (3090 / t4000 / 6000)
  • Seed always-on (nvenc-hevc, cuvslam), slottable (filters, sam2-tiny, clip, klein-4b, qwen-9b if absent marked), never-on-thor (flux-dev, wan-14b, h3, …)
  • Exclusive tags: klein-27b, sam2-maxine-4k
  • Mode presets as arrays of cartridge ids (master-capture, snap, ad, ar, talent-tracking)
  • Envelope ids: 3090 (fractal1 live), t4000 (second scale), 6000
  • Point names at MODELS.md / MODES.md; FACT costs at 3090-SIM.md — do not fork product ground

From advise 2026-09-10 (agt-fable-5-1-arch-review)

  • Spec: pack semantics on a swap envelope (3090 = one slot at a time; max-of-peaks or stated rule). Cite add-sim-box-mix
  • Spec: sam2-maxine-4k vs talent-tracking preset — resolution-variant cart ids, or drop the tag this pass
  • Spec: null scenario — always-on with null cost on target envelope is absent and logged; slottable with null cost cannot seat
  • Record: t4000 cost is unified (unified_gb), or state the sum rule; do not carry a 3090 host/VRAM split onto Thor
  • Record: latency (LIVE/NEAR/MIN from MODELS.md) or pack (SMALL/LARGE from sim-box-mix), or state that this pack is memory-and-power only
  • Record: job per cart pointing at the MODELS.md row (ids are models, MODELS.md rows are jobs)
  • Record: each cost carries src (3090-SIM anchor + date) or fact: false; the design’s klein vram_gb: 16.0 is disk weight size, not a measured VRAM peak
  • Spec note: never-on-thor shelf is an override that wins over any measured cost (policy, e.g. H3)
  • Envelope ceilings (VRAM / host or unified / watts / NVENC sessions) for 3090, t4000, 6000 — owned here or by add-sim-box-mix hardware-kit; one of the two must
  • Acceptance: every seeded preset seats on its home envelope under the spec’s own rules (catches the three spec pins before a packer exists) — watts is not a pack axis; checker is memory + NVENC

From re-advise 2026-09-10 (fable-5-1-arch-review)

  • Spec: power axis — watts is a record and sag (hardware-kit) enforces power, so strike watts as a pack axis from “Latency vs pack” and “Envelope ceilings”
  • Spec: swap formula — Σ(always-on) + max(slottable) against the ceiling, not max over the whole set (always-on carts are resident beside the loaded slot)

From re-advise 2 2026-09-10 (fable-5-1-arch-review)

  • Advise: accept (reviews/2026-09-10-re2-advise.md). Both re-advise sentences closed; acceptance exits 0
  • Fold note: rewrite the three add-sim-box-mix cites in the delta to hardware-kit “Resident 30 fps design target” (change folded at 860bae4); refresh design.md “max-of-peaks” sentence from the spec
  • add-cartridge-pack: checker 4K-pair refusal and 3090 swap-set scenarios in main() (tracked as add-cartridge-pack)